cc @Fidget-Spinner

@Fidget-Spinner: Please review again, I found another bug in _BINARY_OP_INPLACE_ADD_UNICODE opcode. I updated the PR to fix this opcode as well.

            PyUnicode_Append(&temp, right_o);
            Py_DECREF(right_o);
            ERROR_IF(temp == NULL);
            *target_local = PyStackRef_FromPyObjectSteal(temp);

I don't know if *target_local should be set to PyStackRef_NULL on error.

The code in the main branch is different:

            DEAD(left);
            PyUnicode_Append(&temp, right_o);
            _Py_DECREF_SPECIALIZED(right_o, _PyUnicode_ExactDealloc);
            *target_local = PyStackRef_NULL;
            ERROR_IF(temp == NULL);
            res = PyStackRef_FromPyObjectSteal(temp);

I don't know if *target_local should be set to PyStackRef_NULL on error.

I think setting it to PyStackRef_NULL might be wrong? Not too sure. Basically at that point, *target_local refcnt == 1. If we leave it in locals, it must be PyStackref_CLOSE(frame->localsplus[i]) when the frame is cleared during exception raising.

If you want to be sure. Write a test where the append fails (you can trigger an artificial failure somehow), and check if there's reference leaks. If there's refleaks, then we should not be setting it to PyStackRef_NULL and main is wrong.